x-transformers 1.44.2__tar.gz → 1.44.5__tar.gz

{x_transformers-1.44.2/x_transformers.egg-info → x_transformers-1.44.5}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.2
 Name: x-transformers
-Version: 1.44.2
+Version: 1.44.5
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang
@@ -19,3 +19,12 @@ Requires-Dist: einops>=0.8.0
 Requires-Dist: loguru
 Requires-Dist: packaging>=21.0
 Requires-Dist: torch>=2.0
+Dynamic: author
+Dynamic: author-email
+Dynamic: classifier
+Dynamic: description-content-type
+Dynamic: home-page
+Dynamic: keywords
+Dynamic: license
+Dynamic: requires-dist
+Dynamic: summary

{x_transformers-1.44.2 → x_transformers-1.44.5}/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.44.2',
+  version = '1.44.5',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

{x_transformers-1.44.2 → x_transformers-1.44.5}/tests/test_x_transformers.py

@@ -614,7 +614,8 @@ def test_hyper_connections(tanh):
 
     model(x)
 
-def test_hybrid():
+@pytest.mark.parametrize('hybrid_axial_dim', (1, 4))
+def test_hybrid(hybrid_axial_dim):
     from torch.nn import GRU
 
     dec = TransformerWrapper(
@@ -625,6 +626,7 @@ def test_hybrid():
             depth = 6,
             heads = 8,
             attn_dim_head = 64,
+            attn_hybrid_fold_axial_dim = hybrid_axial_dim,
             attn_hybrid_module = GRU(128, 64 * 8, batch_first = True)
         )
     )
@@ -641,6 +643,7 @@ def test_hybrid():
             depth = 6,
             heads = 8,
             attn_dim_head = 64,
+            attn_hybrid_fold_axial_dim = hybrid_axial_dim,
             attn_hybrid_module = GRU(128, 64 * 4, batch_first = True, bidirectional = True)
         )
     )

{x_transformers-1.44.2 → x_transformers-1.44.5}/x_transformers/x_transformers.py

@@ -10,7 +10,7 @@ import torch
 from torch.amp import autocast
 import torch.nn.functional as F
 from torch import nn, einsum, Tensor
-from torch.utils._pytree import tree_flatten
+from torch.utils._pytree import tree_flatten, tree_unflatten
 from torch.nn import Module, ModuleList, ModuleDict
 
 from functools import partial, wraps
@@ -966,6 +966,42 @@ class ShiftTokens(Module):
         x = torch.cat((*segments_to_shift, *rest), dim = -1)
         return self.fn(x, **kwargs)
 
+class FoldAxially(Module):
+    def __init__(
+        self,
+        axial_dim,
+        fn: Module
+    ):
+        super().__init__()
+        self.fn = fn
+        self.axial_dim = axial_dim # will fold the sequence as rearrange("b (n axial_dim) ... -> (b axial_dim) n ...")
+
+    def forward(
+        self,
+        x,
+        **kwargs
+    ):
+        if self.axial_dim == 1:
+            return self.fn(x, **kwargs)
+
+        seq_len, axial_dim = x.shape[1], self.axial_dim
+
+        next_multiple = math.ceil(seq_len / axial_dim) * axial_dim
+        x = pad_at_dim(x, (0, next_multiple - seq_len), dim = 1)
+
+        x = rearrange(x, 'b (n axial_dim) ... -> (b axial_dim) n ...', axial_dim = axial_dim)
+
+        out = self.fn(x, **kwargs)
+
+        (out, *rest_out), tree_spec = tree_flatten(out)
+
+        out = rearrange(out, '(b axial_dim) n ... -> b (n axial_dim) ...', axial_dim = axial_dim)
+
+        out = out[:, :seq_len]
+        out = tree_unflatten((out, *rest_out), tree_spec)
+
+        return out
+
 # post branch operator
 
 class LayerScale(Module):
@@ -1140,6 +1176,7 @@ class Attention(Module):
         custom_attn_fn: Callable | None = None,
         hybrid_module: Module | None = None,
         hybrid_mask_kwarg: str | None = None,
+        hybrid_fold_axial_dim: int | None = None,
         one_kv_head = False,
         kv_heads = None,
         shared_kv = False,
@@ -1341,8 +1378,12 @@ class Attention(Module):
 
         # hybrid module, in same vein as hymba https://www.arxiv.org/abs/2411.13676
 
-        self.hybrid_module = deepcopy(hybrid_module) if exists(hybrid_module) else None
+        hybrid_module = maybe(deepcopy)(hybrid_module)
+
+        if exists(hybrid_module) and exists(hybrid_fold_axial_dim):
+            hybrid_module = FoldAxially(axial_dim = hybrid_fold_axial_dim, fn = hybrid_module)
 
+        self.hybrid_module = hybrid_module
         self.hybrid_mask_kwarg = hybrid_mask_kwarg # for bidirectional, can forward `mask` into the hybrid module and let it handle variable lengths
 
         # output dimension by default same as input, but can be overridden
@@ -2183,7 +2224,7 @@ class AttentionLayers(Module):
         layer_variables = tuple(tuple(layer_variable[i] for i in layers_execute_order) for layer_variable in layer_variables)
 
         # derived input for reinjection if needed
-
+        inp_inject = None
         if self.reinject_input:
             assert not exists(in_attn_cond)
             inp_inject = self.reinject_input_proj(x)
@@ -2241,7 +2282,7 @@ class AttentionLayers(Module):
                 post_branch_norm = maybe(partial)(post_branch_norm, **norm_kwargs)
                 post_main_norm = maybe(partial)(post_main_norm, **norm_kwargs)
 
-            if self.reinject_input:
+            if exists(inp_inject):
                 x = x + inp_inject
 
             if exists(pre_norm):

{x_transformers-1.44.2 → x_transformers-1.44.5/x_transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.2
 Name: x-transformers
-Version: 1.44.2
+Version: 1.44.5
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang
@@ -19,3 +19,12 @@ Requires-Dist: einops>=0.8.0
 Requires-Dist: loguru
 Requires-Dist: packaging>=21.0
 Requires-Dist: torch>=2.0
+Dynamic: author
+Dynamic: author-email
+Dynamic: classifier
+Dynamic: description-content-type
+Dynamic: home-page
+Dynamic: keywords
+Dynamic: license
+Dynamic: requires-dist
+Dynamic: summary